Data Scientist Technical Interview Questions & Answers (2026)

I'd start with EDA: check class distribution (imbalanced?), missing values, feature types (categorical vs numerical), and correlations. With 500 features, I'd first look for obvious drops: remove constant or near-constant features, highly correlated pairs (keep one), and features with >50% missing values. Next, I'd establish a baseline with a simple model — logistic regression or random forest on a sample of 500K rows. This gives a benchmark and helps identify which features matter (feature importance). With 10M rows, I'd check if more data helps or if we've hit diminishing returns by plotting learning curves on subsamples (100K, 500K, 1M, 5M). For feature engineering, I'd create interaction features between the top 20 most important features, handle categoricals with target encoding (careful about leakage — use cross-validated encoding), and consider dimensionality reduction (PCA) on numerical feature groups. For the final model, I'd likely use gradient boosting (XGBoost/LightGBM) — it handles the feature space well, is fast to train on 10M rows, and provides good baseline performance. I'd evaluate with stratified cross-validation using business-relevant metrics (precision-recall, not just AUC, depending on the cost of false positives vs false negatives). Finally, I'd consider inference time constraints for deployment.

Bias measures how far off predictions are from the true values on average — it's systematic error from wrong assumptions. High bias means the model is too simple to capture the real patterns (underfitting). A linear regression trying to fit a quadratic relationship has high bias. Variance measures how much predictions change across different training samples. High variance means the model is too sensitive to the specific training data (overfitting). A deep decision tree that memorizes training data has high variance. The trade-off: as model complexity increases, bias decreases (the model can capture more patterns) but variance increases (it starts fitting noise). The sweet spot minimizes total error = bias² + variance + irreducible noise. Practically, I diagnose this by comparing training error and validation error. Large gap = high variance (try regularization, more data, simpler model). Both errors high = high bias (try more features, complex model, feature engineering). In practice, I usually start simple (high bias, low variance) and gradually increase complexity, monitoring validation error. Ensemble methods like random forests reduce variance by averaging, while boosting reduces bias by iteratively correcting errors. The choice depends on the data size: with little data, I prefer low-variance models; with abundant data, I can afford more complex models.

I'd use DENSE_RANK to handle ties — if two products tie for 2nd place, both get rank 2 and the next product gets rank 3 (unlike RANK which would skip to 4, or ROW_NUMBER which would arbitrarily break the tie). Query: WITH ranked AS (SELECT category, product_id, SUM(revenue) as total_revenue, DENSE_RANK() OVER (PARTITION BY category ORDER BY SUM(revenue) DESC) as rnk FROM orders GROUP BY category, product_id) SELECT category, product_id, total_revenue, rnk FROM ranked WHERE rnk <= 3 ORDER BY category, rnk. The CTE first calculates total revenue per product per category, then ranks within each category. The outer query filters to top 3 ranks. I chose DENSE_RANK because the question says 'including ties' — if two products are tied for 3rd, both should appear. If the interviewer wanted exactly 3 rows per category regardless of ties, I'd use ROW_NUMBER with a tiebreaker (like product_id). I'd also discuss performance: with a large orders table, an index on (category, product_id, revenue) would help the GROUP BY and window function.

I'd evaluate at three levels: offline, online, and business impact. Offline metrics: Precision@K measures what fraction of the top K recommendations are relevant. Recall@K measures what fraction of all relevant items appear in top K. NDCG (Normalized Discounted Cumulative Gain) accounts for ranking position — relevant items ranked higher are scored more favorably. MAP (Mean Average Precision) summarizes precision across all recall levels. I'd also measure catalog coverage (what percentage of items ever get recommended) and diversity (how different recommendations are across users). However, offline metrics often don't correlate perfectly with business outcomes. For online evaluation, I'd use A/B testing measuring CTR, conversion rate, time spent, and crucially, long-term engagement (not just immediate clicks). A model that maximizes CTR might recommend clickbait, not genuinely useful items. I'd also track the distribution of recommendations: are we creating filter bubbles? Business metrics: revenue per user, retention rate, and customer satisfaction scores. The key insight is that the best recommendation system isn't the one with the highest NDCG — it's the one that balances relevance, diversity, freshness, and business goals. I'd also measure novelty (are we recommending things the user wouldn't have found on their own?) as a health metric.

I approach class imbalance at three levels. First, evaluation: never use accuracy for imbalanced datasets. A model predicting 'not fraud' for 99.9% accuracy is useless. I'd use precision-recall AUC, F1 score, or a custom metric based on the business cost of false positives vs false negatives. Second, data-level approaches: oversampling the minority class (SMOTE creates synthetic examples by interpolating between existing minority samples), undersampling the majority class (effective when you have abundant data), or a combination. I'd be careful with SMOTE on high-dimensional data where synthetic examples may not represent real patterns. Third, algorithm-level: use class_weight='balanced' in scikit-learn (adjusts loss function to penalize minority misclassification more), or use algorithms naturally suited for imbalance like gradient boosting with scale_pos_weight. For extreme imbalance (1:10000 like fraud), I've found that combining undersampling with ensemble methods works well — train multiple models on different undersampled majority subsets and ensemble their predictions. The right approach depends on the severity of imbalance and the business cost structure. If missing a fraud case costs $10K but a false alarm costs $10 in manual review, I'd optimize for high recall even at the expense of precision.